Are We More Accurate? Revisiting the European Commission's Macroeconomic Forecasts - EUROPEAN ECONOMY
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
ISSN 2443-8022 (online) Are We More Accurate? Revisiting the European Commission’s Macroeconomic Forecasts Andras Chabin, Sébastien Lamproye and Milan Výškrabka DISCUSSION PAPER 128 | JULY 2020 EUROPEAN ECONOMY EUROPEAN Economic and Financial Affairs
European Economy Discussion Papers are written by the staff of the European Commission’s Directorate-General for Economic and Financial Affairs, or by experts working in association with them, to inform discussion on economic policy and to stimulate debate. DISCLAIMER The views expressed in this document are solely those of the author(s) and do not necessarily represent the official views of the European Commission. This manuscript was completed before the economic impact of the COVID-19 pandemic reached Europe. Authorised for publication by José Eduardo Leandro, Director for Policy, Strategy and Communication. LEGAL NOTICE Neither the European Commission nor any person acting on behalf of the European Commission is responsible for the use that might be made of the information contained in this publication. This paper exists in English only and can be downloaded from https://ec.europa.eu/info/publications/economic-and-financial-affairs-publications_en. Luxembourg: Publications Office of the European Union, 2020 PDF ISBN 978-92-76-11196-2 ISSN 2443-8022 doi:10.2765/220245 KC-BD-19-015-EN-N © European Union, 2020 Non-commercial reproduction is authorised provided the source is acknowledged. For any use or reproduction of material that is not under the EU copyright, permission must be sought directly from the copyright holders. CREDIT Cover photography: © iStock.com/shironosov
European Commission Directorate-General for Economic and Financial Affairs Are We More Accurate? Revisiting the European Commission’s Macroeconomic Forecasts Andras Chabin, Sébastien Lamproye and Milan Výškrabka Abstract In this paper, we present the results of the comprehensive assessment of the accuracy of European Economic Forecasts. High-quality macroeconomic forecasts are a prerequisite for economic surveillance of the European Commission. We evaluate forecasts for three key variables – GDP growth, consumer price inflation and the general government budget balance – on two forecast horizons – current year and one- year-ahead – over the period 2000-2017. Pointing to some improvement in the accuracy recently, the forecasts continue to show a satisfactory track record which does not differ much from the forecast track records of other international institutions. The Commission’s forecasts present largely an unbiased outlook for near term economic developments, accurately foresee an acceleration and deceleration in the underlying variables and mostly contain information beyond a naïve forecast. There is room for improvement, however. The forecasts appear to be prone to repeating errors, which to some extent seems to be related to an overly conservative assessment of the business cycle dynamics and to a lesser extent to errors in technical assumptions. JEL Classification: C1, E60, E66. Keywords: macroeconomic forecasts, forecast errors, accuracy, statistical properties, GDP, inflation, government budget balance, European Commission. Acknowledgements: We would like to thank Susanne Hoffmann, João Miguel Leal, Tsvetan Tsalinski, Evelyne Hespel and Björn Döhring for reviewing and discussing earlier versions of the paper. All remaining errors are solely the responsibility of the authors. Contact: Milan Výškrabka, European Commission, Directorate-General for Economic and Financial Affairs, milan.vyskrabka@ec.europa.eu. Andras Chabin and Sébastien Lamproye prepared the paper while working at the European Commission, Directorate-General for Economic and Financial Affairs. EUROPEAN ECONOMY Discussion Paper 128
CONTENTS 1. Introduction ........................................................................................................................................ 7 2. Data description and summary statistics ..................................................................................... 8 2.1. Definitions and reference periods ......................................................................................................8 2.2. Summary statistics ................................................................................................................................10 2.2.1. Mean error.................................................................................................................................................. 10 2.2.2. Mean absolute error ................................................................................................................................. 10 2.2.3. Root mean squared error ........................................................................................................................ 10 3. Forecasting performance in recent years ................................................................................. 11 3.1. Gross domestic product .....................................................................................................................11 3.2. Inflation ..................................................................................................................................................12 3.3. General Government balance .........................................................................................................13 3.4. Volatility adjusted forecast errors......................................................................................................14 4. Properties of the forecast errors ................................................................................................... 15 4.1. Are the projections biased? ..............................................................................................................15 4.1.1. Gross Domestic Product........................................................................................................................... 16 4.1.2. Inflation ....................................................................................................................................................... 16 4.1.3. General government balance ............................................................................................................... 17 4.2. Are the projection errors persistent? ................................................................................................18 4.2.1. Gross Domestic Product........................................................................................................................... 18 4.2.2. Inflation ....................................................................................................................................................... 18 4.2.3. General government balance ............................................................................................................... 19 4.3. Are the projections quantitatively accurate? ................................................................................20 4.3.1. Gross Domestic Product........................................................................................................................... 21 4.3.2. Inflation ....................................................................................................................................................... 21 4.3.3. General government balance ............................................................................................................... 21 4.4. Do the projections encompass the naïve forecast?.....................................................................22 4.4.1. Gross Domestic Product........................................................................................................................... 23 4.4.2. Inflation ....................................................................................................................................................... 23 4.4.3. General government balance ............................................................................................................... 24 3
4.5. Are the projections directionally accurate? ..................................................................................24 4.5.1. Gross Domestic Product........................................................................................................................... 24 4.5.2. Inflation ....................................................................................................................................................... 24 4.5.3. General government balance ............................................................................................................... 25 4.6. Are the projections efficient? ............................................................................................................25 4.6.1. Gross Domestic Product........................................................................................................................... 26 4.6.2. Inflation ....................................................................................................................................................... 26 4.6.3. General government balance ............................................................................................................... 26 5. The role of external Factors ........................................................................................................... 27 5.1. The role of external assumptions .......................................................................................................27 5.1.1. GDP growth forecasts .............................................................................................................................. 28 5.1.2. Inflation forecasts ...................................................................................................................................... 29 5.2. Cyclicality of forecast errors ..............................................................................................................29 6. Comparing GDP forecast errors with those of other international institutions .................... 32 6.1. Commission versus OECD ...................................................................................................................32 6.2. Commission versus IMF ........................................................................................................................33 6.3. Commission versus Consensus...........................................................................................................34 6.4. Commission versus ECB.......................................................................................................................35 7. Conclusions ...................................................................................................................................... 35 LIST OF TABLES Table 2.1.: Number of observations in reference periods ............................................................................................................ 9 Table 4.1.: Tests for forecast bias, 2000-2017 ................................................................................................................................ 17 Table 4.2.: Tests for error persistence, current-year, 2000-2017 ................................................................................................. 19 Table 4.3.: Tests for persistence, year-ahead, 2000-2017 ........................................................................................................... 20 Table 4.4.: Diebold-Mariano test, 2000-2017 ................................................................................................................................ 22 Table 4.5.: Forecast encompassing tests ...................................................................................................................................... 23 Table 4.6.: Tests for directional accuracy, 2000-2017 ................................................................................................................. 25 Table 4.7.: Forecast efficiency tests ............................................................................................................................................... 26 Table 5.1.: External assumptions, gross domestic product, current year, 2000-2017 ............................................................. 28 Table 5.2.: External assumptions, inflation, current year, 2000-2017 ......................................................................................... 29 4
Table 5.3.: Forecast efficiency tests ............................................................................................................................................... 31 Table A.1: Gross domestic product, error statistics, 2000-2017 .................................................................................................. 38 Table A.2: Inflation, error statistics, 2000-2017 .............................................................................................................................. 39 Table A.3: General government budget balance, error statistics, 2000-2017 ........................................................................ 40 Table A.4: Tests for forecast bias, 1969-2017 ................................................................................................................................ 41 Table A.5: Tests for error persistence, current-year, 1969-2017 .................................................................................................. 42 Table A.6: Tests for error persistence, year-ahead, 1969-2017 .................................................................................................. 43 Table A.7: Diebold-Mariano tests, 1969-2017 ............................................................................................................................... 44 Table A.8: Tests for directional accuracy, 1969-2017.................................................................................................................. 45 Table A.9: External assumptions, gross domestic product, year-ahead, 2000-2017 .............................................................. 45 Table A.10: External assumptions, inflation, year-ahead, 2000-2017 ........................................................................................ 46 LIST OF GRAPHS Graph 3.1.: Gross domestic product, absolute error .................................................................................................................. 11 Graph 3.2.: Inflation, absolute error ............................................................................................................................................... 13 Graph 3.3.: General government balance, absolute error ....................................................................................................... 14 Graph 3.4.: Volatility adjusted forecast errors, 2000-2017, RMSE, current-year ...................................................................... 15 Graph 5.1.: GDP growth, data and current year forecast errors, 2000-2017 .......................................................................... 30 Graph 5.2.: Inflation, data and current year forecast errors, 2000-2017 ................................................................................. 31 Graph 6.1.: Mean absolute error, 2000-2017 ................................................................................................................................ 33 Graph 6.2.: Mean absolute error, 2000-2017 ................................................................................................................................ 33 Graph 6.3.: Mean absolute error, 2000-2017 ................................................................................................................................ 34 Graph 6.4.: Mean absolute error, 2000-2017 ................................................................................................................................ 35 Graph A.1: GDP growth, data and year-ahead forecast errors, 2000-2017 .......................................................................... 46 Graph A.2: Inflation, data and year-ahead forecast errors, 2000-2017 .................................................................................. 47 Graph A.3: Mean absolute error, EC and OECD forecasts, 1969-2017.................................................................................... 47 Graph A.4: Mean absolute error, EC and IMF forecasts, 1969-2017 ........................................................................................ 48 Graph A.5: Mean absolute error, EC and Consensus Economics forecasts, 1969-2017 ....................................................... 48 REFERENCES ANNEX 5
1. INTRODUCTION The European Economic Forecasts are an integral part of the European Commission’s Treaty-based economic surveillance framework. 1 In particular, the forecasts, prepared by the staff of the European Commission, form a basis for fiscal surveillance, the surveillance of macroeconomic imbalances and the formulation of economic policy recommendations to the euro area and the Member States. 2 However, the importance of macroeconomic forecasting goes well beyond its role in the Commission’s surveillance. Policy institutions, including the European Commission, employ macroeconomic forecasts also to communicate their views on economic prospects, shape the basis for policy considerations and steer expectations. In this broader application, the narrative and the assessment of risks are also key attributes of the forecast. Qualitative consistency is thus equally important and the forecast process puts strong emphasis not only on quantitative accuracy but also on overall consistency across variables and Member States. In this paper, we focus on the quantitative accuracy of forecasts instead. We evaluate point estimates of three prominent variables in the Commission’s economic surveillance, i.e. GDP growth, inflation and the general government budget balance. Each spring and autumn, an evaluation of Member States fiscal plans based on the Commission’s macroeconomic forecasts widely informs the overall assessment of compliance with the EU fiscal rules. For this reason, quality forecasts are a necessary ingredient of the Commission’s surveillance framework. A macroeconomic forecast is inherently an imprecise description of actual future economic developments. The role of a forecaster is to accurately assess the current state of the economy and take all available information into account when producing a forecast. To increase the transparency and credibility of its forecasts, the Commission regularly evaluates the forecast performance. Fioramanti et al. (2016) extensively discussed the results of the last such exercise. They find that “the European Commission’s forecasts continue to display a reasonable track record, similar to that of the other international institutions” over the entire forecast history until 2014. 3 Using a battery of statistical tests, the forecasts were found unbiased for most EU Member States, and adequately anticipated pick-ups and slow-downs. They also were more accurate than simple naïve forecasts. Moreover, the forecast errors were mostly random. The results mostly confirmed the findings of previous Commission’s assessments of its forecast performance (Keereman (1999), Melander, Sismanidis and Grenouilleau (2007), González Cabanillas and Terzi (2012)). Based on a similar set of tests, forecasters at other international institutions, such as the European Central Bank, the International Monetary Fund and the OECD, also regularly carry out the assessment of the accuracy of their forecasts. Kontogeorgos and Lambrias (2019) concluded that (Broad) Macroeconomic Projection Exercises of the ECB reliably informed policymakers about future economic developments in the euro area. Yet, they also found room for improvement. Pain et al. (2014) compared the performance of OECD forecasts during and after the financial crisis. They found that projections mostly overestimated GDP growth, failed to anticipate the magnitude of the impact of the crisis and the subsequent lukewarm recovery. Similarly, Eicher et al. (2018) found that IMF forecasts for countries in times of crisis carried informational value. Bias and inefficiency were present in projections of some variables for some countries. In this iteration of the Commission’s evaluation, we extend the database by adding three years of forecasts, 2015-2017. During this period, the European economy experienced a sustained expansion 1 Hereinafter referred to as the Commission’s forecasts or simply the forecasts. 2 This analysis evaluates Commission’s forecasts until 2017, hence the UK is included and is also part of the EU aggregates calculated throughout the paper. As of 1 February 2020, the UK is no longer a Member State of the EU. During the transition period, the UK remains subject to economic and fiscal surveillance. 3 Starting in 1969 for some Member States. 7
with most Member States’ growth accelerating. The upswing of the business cycle did not go hand in hand with rising inflationary pressures, however. By contrast, the environment of low interest rates and swift growth helped most Member States reduce (increase) their budget deficits (surplus). Compared to earlier years, when the 2008-2009 financial crisis and the sovereign debt crisis severely challenged the performance of forecasts, the newly added period to the sample was characterised by much less volatile developments. The first aim of this paper is to assess the relative performance of recent forecasts compared to findings in the previous evaluation exercise. To this end, we calculate basic statistics widely used to characterise forecast accuracy, namely the average error, the mean average error and the root mean square error. To explore some sources of inaccuracies, we estimate the contribution of errors in external assumptions. The second aim is to test the quality of the forecasts formally. Quality forecasts are required to be unbiased, efficiently use the information available and not repeat past errors. The third aim is to compare the Commission’s forecasts with those of the OECD, IMF, the ECB and Consensus Economics. The paper is organised as follows: section 2 introduces the data and the summary statistics used in the analysis. Section 3 explores the quantitative accuracy of the forecasts and discusses the performance of the newly added forecasts to the sample. Section 4 presents the qualitative characteristics of the forecasts based on the results of several formal statistical tests. Section 5 investigates the role of several factors that are likely to drive forecast inaccuracies. Section 6 compares the accuracy of the Commission’s forecasts with forecasts of other major international institutions and private forecasters. Finally, we conclude. 2. DATA DESCRIPTION AND SUMMARY STATISTICS 2.1. DEFINITIONS AND REFERENCE PERIODS Three prominent forecast variables are considered in this paper: real GDP growth, inflation (price deflator of private consumption) and the general government balance to GDP ratio. These variables are essential to economic analysis and policy debate. This paper deals with the current year forecast error and the year-ahead forecast error. The forecast error for a given country i is defined as follows: where yi,t,t and yi,t+1,t are the forecasts made for country i at time t, for periods t and t+1 respectively; yi,t and yi,t+1 are the realisations of the variable in question for country i for period t and t+1, respectively. Hence, positive errors indicate an overestimation, whereas negative errors indicate an underestimation of the actual outturns. Data have been processed in a similar manner as in previous evaluations of the Commission forecasts’ accuracy. 4 The current year forecasts (yi,t,t) and current year realisations (yi,t) are extracted from the Commission's spring forecast publications, which are published in May. The current year forecast for period t is taken from the spring forecast in period t, while the realisation for period t is taken from the spring forecast in period t+1. The year-ahead forecasts (yi,t+1,t) and realisations (yi,t+1) are taken from the Commission's autumn forecasts, which are published in November. The year-ahead forecasts for 4 Keereman (1999), Melander, Sismanidis and Grenouilleau (2007), González Cabanillas and Terzi (2012), Fioramanti et al. (2016) 8
period t are taken from the autumn forecast in t-1, while the realisation of the forecast for period t is taken from the autumn forecast t+1. The forecast errors are calculated for all EU Member States except for Croatia, the UK and for the euro area and European Union (EU) aggregates. 5 For the EU and the euro area, the aggregates reflect the changing composition over time. We focus mainly on the forecast performance in recent history, starting in 2000 and ending in 2017. This period is referred to as the baseline sample. For the sake of continuity, we include the results of the previous study in the next section. Wherever appropriate, we complement the empirical analysis with statistical tests run on the entire available sample, starting in 1969 for some Member States, in order to verify empirical findings based on the relatively short baseline sample. Table 2.1.: Number of observations in reference periods General GDP Inflation Government Balance '69-'17 '00-'17 '69-'17 '00-'17 '69-'17 '00-'17 Belgium 49 18 49 18 47 18 Germany 49 18 49 18 49 18 Estonia 14 14 14 14 14 14 Ireland 45 18 45 18 44 18 Greece 37 18 37 18 36 18 Spain 32 18 32 18 32 18 France 49 18 49 18 49 18 Italy 49 18 49 18 49 18 Cyprus 14 14 14 14 14 14 Latvia 14 14 14 14 14 14 Lithuania 14 14 14 14 14 14 Luxembourg 49 18 49 18 44 18 Malta 14 14 14 14 14 14 Netherlands 49 18 49 18 49 18 Austria 23 18 23 18 23 18 Portugal 32 18 32 18 32 18 Slovenia 14 14 14 14 14 14 Slovakia 14 14 14 14 14 14 Finland 23 18 23 18 23 18 Euro area 20 18 20 18 20 18 Bulgaria 11 11 11 11 11 11 Czech Republic 14 14 14 14 14 14 Denmark 45 18 45 18 41 18 Hungary 14 14 14 14 14 14 Poland 14 14 14 14 14 14 Romania 11 11 11 11 11 11 Sweden 23 18 23 18 23 18 United Kingdom 45 18 45 18 45 18 EU 49 18 49 18 49 18 Source: EC, Eurostat, own calculations. 5 Croatia is left out due to an insufficient number of observations. 9
2.2. SUMMARY STATISTICS To measure the European Commission's forecasting performance, several summary statistics are computed in this study. They are described in the paragraphs below. 2.2.1. Mean error The mean error (ME) is the average forecast error for each country i over a given period T. It is a basic indicator of accuracy. This indicator indicates possible bias in the forecast as positive and negative errors can offset each other. Formally, for the current year and the year-ahead forecast, respectively. 2.2.2. Mean absolute error The mean absolute error (MAE) is the average of the absolute value of the forecast errors for each country i over a given period T. Negative errors cannot cancel out positive ones, therefore MAE does not limit the size of the error. The MAE, however, does not provide information on the direction of the error (underestimation or overestimation). Also, in the calculation of the MAE, all the errors are equally weighted in the average whatever their size. Formally, 1 = �� , , � =1 1 = �� , +1, � =1 for the current year and the year-ahead forecast, respectively. 2.2.3. Root mean squared error The root mean squared error (RMSE) is the square root of the average of the squared forecast errors for each country i over a given period T. Since the errors are squared, large errors have a relatively higher weight. Therefore, the RMSE is preferred when large errors are considered particularly harmful. It is an estimate of the standard deviation of the error series. Also, the RMSE is not independent of the number of observations and does not provide information on the direction of the errors. Formally, for the current year and the year-ahead forecast, respectively. 10
3. FORECASTING PERFORMANCE IN RECENT YEARS In this section, we discuss the performance of recent macroeconomic forecasts by comparing errors over the period 2000-20017 (baseline sample) to the reference period 2000-2014. This section also subjects the forecasts to standard metrics of forecast errors that provide. We discuss the ME, MAE and RMSE for real GDP growth, inflation and general government balance in percentage of GDP. Detailed summary statistics are provided in Annex. 3.1. GROSS DOMESTIC PRODUCT The current year forecasts of GDP growth performed reasonably well in the years 2015-2017. Adding these years to the reference sample 2000-2014 slightly improves the overall forecast accuracy. As measured by both the MAE and the RMSE, the forecast error for real GDP growth for the current year declined somewhat for both the EU and euro area aggregates (Table A.1 in Annex). Graph 3.1 shows that the forecasts undershot the economic expansion in 2015-2017 after having overestimated the expansion in the preceding years. Hence the average error (ME) decreased more significantly compared to the other two measures of accuracy. Further back, forecast errors for the current year were particularly high in the years 2001 (bursting of the dotcom bubble), 2008 (the financial crisis) and 2010 (sovereign debt crisis); all dominated by major shocks. The forecast also did not accurately anticipate the strength of growth in 2017. The current year forecast errors for both the EU and the euro area for the crisis and post-crisis period (2008-2017) are slightly lower compared to those for the pre- crisis period (2000 to 2007). Compared to historical standards, the current year growth forecasts fared particularly well in the period between 2011 and 2016. Graph 3.1.: Gross domestic product, absolute error Current year Year-ahead 1.4 5.0 pps. pps. 4.5 1.2 4.0 1.0 3.5 3.0 0.8 2.5 0.6 2.0 1.5 0.4 1.0 0.2 0.5 0.0 0.0 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 EU EA Av. EU Av. EA EU EA Av. EA Av. EA Source: EC, Eurostat, own calculations. For the year-ahead forecasts, errors are markedly larger than for the current year forecasts as much less information is available at the time of forecasting. Similarly to the current year forecasts, the overall accuracy improved for both the EU and euro area aggregates in 2015-2017. The MAE and RMSE fell slightly while the ME decreased more significantly. The year-ahead forecast for 2009 projected some slowdown in economic activity in both the EU and the euro area, but failed to anticipate the contraction. The forecast error for 2009 is by far the largest error in the sample period. By contrast, for the current year forecasts, the forecast error for 2009 was small, as the economy was already formally in recession and forecasters had more information at the time of forecasting. For year-ahead forecasts, the accuracy deteriorated in the crisis and post-crisis period compared to the pre-crisis period for both the EU and euro area aggregates as volatility of GDP growth increased in this period and the forecasts struggled to anticipate factors driving the economy. 11
At the Member State level, the current year forecasts for the years 2015-2017 reduced the overall error in 18 Member States when measured by the ME and 19 Member States when measured by the RMSE. The average reduction in the ME reached about 9% while the RMSE decreased by about 6% on average. Current year forecasts for GDP growth in Latvia, Slovakia, Hungary, the UK and Belgium improved by more than 10%, as measured by the ME. On the other hand, the forecast accuracy worsened in 8 Member States when measured by the ME and in 5 Member States when measured by the RMSE. The average increase in the absolute error was about 10%. More country forecasts recorded an increase in the average error. The increase is either negligible or the error is quantitatively small in most cases, however. For the year-ahead forecast, the results show an improvement in the forecast accuracy in 23 Member States when measured by the ME and 25 Member States when measured by the RMSE. The average decrease in both the mean absolute error and the root mean square error is about 9%. The accuracy of the forecasts for Estonia, Latvia, Lithuania and Slovakia improved more than the accuracy for other countries. The accuracy of the forecasts for Ireland, Greece, Malta and Bulgaria decreased when the forecasts for Ireland recorded a significant decline in overall performance. 6 Similarly, four country forecasts recorded an increase in the mean error – Ireland, Malta, Slovakia and Poland – when Ireland clearly stands out while deterioration for Slovakia and Poland are not substantial. Unlike for the euro area and the EU, some deterioration in the forecast accuracy can be seen for most EU Member States for both the current year and year-ahead forecasts in the post-crisis period compared to the pre-crisis period. The deterioration is, in general, much more pronounced for the year- ahead forecasts. For the current year forecasts, the largest increase in the MAE in the crisis (and post- crisis) period compared to the pre-crisis period can be seen for Cyprus, Greece and Ireland; three countries that were under an Adjustment Programme. There are however some exceptions where the forecast accuracy improved for the current year (e.g. France, the Netherlands, Belgium and Slovakia, Czech Republic and Poland). For the year-ahead forecasts, the same Programme countries reported the largest increases in the MAE. Poland, Slovakia and Belgium stand out as the countries for which the forecast error decreased in the post-crisis period compared to the pre-crisis period. 3.2. INFLATION Errors in the forecasts of inflation measured as the annual rate of change in the price deflator of private consumption are lower than those in GDP growth as the latter is somewhat more volatile. The addition of three years to the sample marginally reduces the MAE and RMSE for both the EU and the euro area. The current year forecasts of inflation for the years 2014-2017 missed the outturns for the euro area and the EU by a narrow margin when in all instances the forecasts overshot actual inflation. Forecast performance in these three years ended a period of substantially larger errors between 2009 and 2014 (see Graph 3.2). In this period, a sharp drop in inflation in 2009 was followed by an equally fast rebound the next year with a gradual slowdown afterwards. The forecasts did not fully envisage the erratic behaviour at the beginning of the period, neither did they recognise the changing trend in the second half of the period. The average error over the whole studied sample is, however, very close to zero. The accuracy of the one year-ahead forecasts slightly improved as well. Both the MAE and the RMSE decreased in both the euro area and the EU by adding three new observations. The magnitude of errors does not differ significantly from the magnitude of errors in a few previous years. Similarly to the current year forecasts, most year-ahead forecasts overestimated the actual outturns since 2013. Despite a series of positive errors, the average error over the whole sample remains very close to zero. 6 The Irish economy is a good example of a challenge that forecasters may face. One-off accounting operations of large multinational corporations, such as those in 2015, can substantially impact overall economic growth with detrimental consequences for the accuracy of any forecast. 12
Graph 3.2.: Inflation, absolute error Current year Year-ahead 0.7 2.5 pps. pps. 0.6 2.0 0.5 1.5 0.4 0.3 1.0 0.2 0.5 0.1 0.0 0.0 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 EU EA Av. EU Av. EA EU EA Av. EU Av. EA Source: EC, Eurostat, own calculations. In general, the accuracy of inflation forecasts deteriorated after the crisis as evidenced at the aggregate (see Graph 3.2) and Member States levels. The magnitude of both current year and one year-ahead forecast errors is substantially larger after 2009. Unlike in the case of the GDP forecasts, the crisis marked the beginning of a prolonged period of reduced accuracy of inflation forecasts. On average, the errors are close to zero in both the pre-crises and post-crisis period, however. At the Member State level, improving forecast records for the current year are also evident for Member States, as the MAE and the RMSE of the current-year forecasts of all Member States decreased as a result of the addition of the 2015-2017 period. With respect to the year-ahead forecasts, the accuracy increased for all Member States except Luxembourg and Denmark with these two countries recording only marginal deterioration in both the MAE and the RMSE. 3.3. GENERAL GOVERNMENT BALANCE Turning to the general government balance-to-GDP ratio, errors in current-year forecasts were larger around the turn of the century than during the 2008-2009 crisis, as was also the case of the current- year GDP forecast errors. The addition of three years of data leads to a small improvement in the forecasting performance in terms of both the MAE (see Table A.3 in Appendix) and RMSE. In all three years, the Commission’s current year forecasts underestimated the pace of fiscal consolidation in both the EU and the euro area. Similarly to the GDP growth forecasts for this period, errors are larger towards the end of the sample as the business cycle culminated. In terms of the average error, the current year forecasts of the general government balance-to-GDP ratio performed well, see Graph 3.3. With the exception of the year 2009, the pattern of the one-year-ahead forecast errors closely resembles the pattern of the current year forecast errors. Especially due to the year 2009, the accuracy statistics of the year-ahead forecasts are worse than those of the current year forecasts. Three newly added observations to the sample are negative, their magnitude is smaller compared to the past errors, which improve overall forecast performance when measured by any performance statistics. Gains in the accuracy are larger when measured by the average error (ME) as the negative errors in 2015-2017 compensate for the overall positive error accumulated in the period before. 13
Graph 3.3.: General government balance, absolute error Current year Year-ahead 1.8 4.5 pps. pps. 1.6 4.0 1.4 3.5 1.2 3.0 1.0 2.5 0.8 2.0 0.6 1.5 0.4 1.0 0.2 0.5 0.0 0.0 00 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 EU EA Av. EU Av. EA EU EA Av. EU Av. EA Source: EC, Eurostat, own calculations. At the Member State level, the accuracy of the current year forecasts improved for 17 Member States when measured by the ME and for 21 Member States when measured by the RMSE. The average improvement in the forecast accuracy is about 10%. On the other hand, the addition of three new observations led to some deterioration in the forecast accuracy in 9 Member States when measured by the MAE and in 5 Member States when measured by the RMSE. The average rate of deterioration is about 7%. In terms of the mean error, the forecast accuracy improved in only 14 Member States while it deteriorated in 11 Member States. In all but two Member States, the current year forecast errors in the period 2014-2017 were negative. The accuracy of the one year-ahead forecasts improved for a majority of the Member States. Only 5 Member States recorded some deterioration in the forecast accuracy; the forecasts for Malta stands out in this group with the largest rise in the ME. All other Member States recorded an improvement in the forecast accuracy. The average gain in improvement reached 10% when measured by the MAE and 8% when measured by the RMSE. 3.4. VOLATILITY ADJUSTED FORECAST ERRORS In general, forecasts for some countries are more accurate than for others irrespective of the variable in question and the forecast horizon as we document in Tables A.1-A.3. There is a number of factors out of forecasters’ control, which may substantially affect the ex post evaluation of forecast performance. Namely, the stability of the economy, changes in accounting standards and data quality with frequent and large revisions are typical factors that make forecasting for some countries more challenging than for others. Naturally, forecasts of more volatile variables are likely to be less accurate in terms of either the MAE or the RMSE. In this section, we explore to what extent the RMSEs are related to the volatility of the underlying data. By normalising the RMSEs by the corresponding standard errors, one can isolate the impact of one determinant of forecast errors, see Graph 3.4. Taking into account the volatility of the underlying data, current year forecasts of GDP growth are more accurate on average than forecasts of the other two variables although only slightly so with respect to the forecasts of inflation. Similarly, the dispersion of relative RMSEs is lowest for GDP growth forecast and largest for the general government balance. In the previous sections, forecasts for the euro area and the EU came out among the most accurate. Unsurprisingly, forecasts for the two areas continue to dominate in terms of relative performance. The fact that forecasts for the two blocks are the result of aggregation of individual forecasts of Member States contributes to higher accuracy as errors of the opposite sign offset one another in aggregation. 14
Graph 3.4.: Volatility adjusted forecast errors, 2000-2017, RMSE, current-year GDP growth Inflation General government balance 1.0 1.0 1.0 EE 0.9 0.9 0.9 0.8 0.8 MT 0.8 EL PL SI CZ BG 0.7 0.7 UK AT 0.7 LU SE IE LU 0.6 MT 0.6 FR SE 0.6 SE MT RO BG DK CY DK IE FI DK NL 0.5 PT CY 0.5 DE LU SI LV 0.5 DE PT AT EE BG CY SK LT BE FR SI BE NL RO IT LV 0.4 IE 0.4 LV HU 0.4 DE AT ES FI CZ ES NL UK CZ PL EU FI IT HU RO PT LT BE EA 0.3 EA 0.3 0.3 SK EL IT EA EE FR EL EU EU SK HU UK LT 0.2 0.2 ES 0.2 0.1 0.1 0.1 0.0 0.0 0.0 relative RMSE EU average relative RMSE EU average relative RMSE EU average Note: RMSE for each Member State is normalised by the respective standard deviation of the data. Source: EC, Eurostat, own calculations. RMSEs are closely related to the standard deviation of the underlying data series. Of course, the volatility in the underlying data is only one of many factors affecting the forecast accuracy, and relative performance of forecasts differ across Member States due to a number of factors as briefly outlined in Introduction. 4. PROPERTIES OF THE FORECAST ERRORS This section focuses on the statistical properties of the forecast errors, as presented in the preceding chapter. As we argue above, a proper assessment of the state of the economy and the use of all relevant information should lead to errors that are random and not large on average. Following the literature, these essential criteria translate into a series of statistical tests detecting the presence of bias, error persistence and informational efficiency, and also challenge the forecasts against a simple benchmark model. 7 Generally, the tests are performed using linear regressions for each of the 27 Member States and for the EU and euro area aggregates where feasible. The feasibility depends mainly on the sample size (see Table 2.1). When the sample size is too small to perform a reliable regression analysis, non- parametric alternatives are used. In case non-parametric tests are not suitable or not available, panel regressions are used to get an overall picture across Member States. In the baseline scenario, the tests are performed on the period 2000-2017. In order to assess the power of statistical tests and stability of regression results, the analysis is complemented by tests run on the entire sample, starting in 1969 for some Member States. The results of the tests performed on the entire sample are reported in Appendix. 4.1. ARE THE PROJECTIONS BIASED? Forecasts from public national or international institutions are often assumed to be too optimistic. Criticism was notably voiced against the Commission's forecast during the Great Recession of 2008- 2009 and the subsequent period of a slower-than-expected recovery. 8 The issue at stake is the presence of bias. Unbiasedness requires the forecast errors to be close to zero on average over the sample. In other words, it prescribes that there is no systematic over- or under-estimation of a selected variable. 7 For example Fioramanti et al. (2016), Bank of England (2015), Kontogeorgos and Lambrias (2019) 8 https://www.euractiv.com/section/social-europe-jobs/news/commission-too-optimistic-in-economic-growth-forecasts- analyst-says/ 15
In order to test whether the Commission's forecasts are biased, the projection errors are regressed on a constant as follows: , , = + , , (1) , +1, = + , +1, (2) Where , , and , +1, stand for the current year and year-ahead forecast errors for country at time respectively, and ε for an independently and identically distributed error term. In the absence of bias, the constant term should not be statistically different from zero, i.e. = 0. Below, an estimate is considered statistically significant when the corresponding p-value is less than 0.05. The null hypothesis of the absence of a bias is then rejected at the 5% significance level. 4.1.1. Gross Domestic Product Fioramanti et al. (2016) found, in line with the previous Commission’s studies, that there was no evidence of bias in the Commission’s projections for GDP growth for the EU and euro area aggregates, except for the year ahead forecasts with respect to the EU when tested on the entire sample. In the updated sample, from 2000 to 2017, the results still hold (Table 4.1). The single-series regressions for both areas indicate a positive but statistically not significant average error in the year- ahead forecasts. Taking into account all Member States data, the panel regression confirms a slight upward bias of the year ahead forecasts. This more powerful tool shows that the positive bias in the year-ahead forecasts is statistically significant at the 1% level. At the Member State level, the forecast for Italy remains the only country forecast to have a bias that is statistically significant at the 1% level, for both the current year and year-ahead forecasts. Compared to the previous Commission study, the existence of bias has been detected for more Member States when extending the observation sample to 2017. With regards to the current year, negative and statistically significant bias was found for the forecasts of Malta and Denmark. For the year ahead forecasts, positive bias was found for the forecasts for Belgium, France, Italy and Portugal. As indicated by the single-country regression results for both areas and the panel regression, current year forecast average error for most Member States is not only quantitatively small, there is also no systematic qualitative pattern. The average error is positive for twelve countries and negative for fifteen countries. The results are markedly different for the year-ahead forecasts. In general, forecasts overestimated GDP growth one-year-ahead when forecasts for only four Member States recorded a negative average error. Member States whose economic growth was underestimated grew swiftly over the period of evaluation (Poland, Slovakia, Malta and Ireland). In general, forecasts underestimated one-year-ahead GDP growth in the pre-crisis years (2000-2007) while in post-crises years, the average error was positive in most countries. The financial crisis and later the sovereign debt crisis in Europe was a challenging period for forecasters when GDP growth persistently underperformed average growth rates. Performance of the forecasts in more recent years (2015-2017), however, helped to reduce the inaccuracy for most Member States. 9 4.1.2. Inflation Unlike the forecasts of GDP growth, Fioramanti et al. (2016) found the Commission’s forecast of inflation for the euro area and the EU unbiased on both forecast horizons. In the updated sample, 2000-2017, the inflation forecasts for the two aggregates continue to display no bias. In the 2000-2017 period, the panel regression shows no bias for both current year and year-ahead inflation forecasts. This also holds for the aggregated euro area and EU variables. However, among Member States, in addition to the bias for Spain (which is now only significant at the 10% level) and 9 see Section 3.1 16
Malta, forecasts for Belgium, France and Austria also came out with a large and statistically significant bias for the current year forecast. In the pre-crisis period (2000-2007), the year-ahead inflation was on average underestimated across for the EU and the euro area (-0.1 pps. and -0.23 pps., respectively), while in the 2008-2017 period, inflation projections were somewhat biased upwards at both horizons. These findings are supported by the panel regressions as well, with coefficient estimates of -0.27 and -0.25 in the pre-crisis period for the current year and year-ahead projections, respectively, and 0.11 and 0.22 in the post-crisis sample for the two horizons. All of the panel regression estimates are significant at the 5% significance level. Table 4.1.: Tests for forecast bias, 2000-2017 General GDP Inflation Government current year current year current year year ahead year ahead year ahead Belgium 0.05 0.31** -0.25*** -0.25*** -0.03 0.00 Germany -0.10 0.23 0.03 0.13** -0.50* -0.39 Estonia -0.06 0.73 -0.23 -0.07 -0.70 -0.94 Ireland -0.71 -1.34 0.11 0.38 1.25 1.35 Greece 0.33 1.21 -0.17* -0.14 1.73** 2.34*** Spain -0.14 0.28 -0.21* -0.25 0.63 0.93 France 0.18* 0.50*** 0.18*** 0.03 0.09 0.19 Italy 0.41*** 0.99*** 0.02 -0.04 0.15** 0.29** Cyprus -0.58 0.80 0.40 0.46 0.04 -0.38 Latvia -0.35 0.48 -1.28 -1.48* -0.55 -0.72 Lithuania 0.05 0.48 -0.26 -0.37 -0.28 -0.50 Luxembourg -0.04 0.29 -0.12 0.06 -0.98*** -1.64*** Malta -0.79** -0.65 0.37** 0.38*** -0.42 -0.70 Netherlands 0.19 0.50 0.10 0.25 -0.30 -0.10 Austria 0.01 0.35* -0.19*** 0.04 -0.43*** -0.39*** Portugal 0.11 0.58** -0.13 0.08 0.27*** 0.98*** Slovenia -0.14 0.50 0.06 0.54 0.43 0.49 Slovakia -0.47 -0.13 0.12 0.36 0.05 0.19 Finland 0.26 0.67 0.01 -0.03 -0.36* -0.32 Euro Area 0.09 0.38 0.02 0.01 -0.01 0.13 Bulgaria -0.05 0.59 0.06 0.41 -0.07 0.70 Czech Republic -0.35 0.13 0.16 0.30 -0.86*** -0.99* Denmark 0.46** 0.76* 0.11 0.09 -0.85*** -0.92** Hungary 0.09 0.46 0.06 0.15 -0.16 -0.59 Poland -0.41* -0.26 0.04 0.18 0.48 0.50 Romania -0.16 0.82 0.31* -0.06 0.16 0.36 Sweden -0.02 0.18 -0.11 0.26** -0.72*** -0.79*** United Kingdom 0.19* 0.39 -0.01 -0.06 -0.14 0.19 EU 0.10 0.34 0.02 0.03 -0.06 0.11 Overall -0.04 0.36*** -0.03 0.04 -0.07 -0.01 Notes: Significance levels: (*) 0.10, (**) 0.05, (***) 0.01. Reported values are the estimated average errors, α, from equations (1) and (2). Source: EC, Eurostat, own calculations. 4.1.3. General government balance In the case of the general government balance, only the budgetary measures known in sufficient detail or approved by the national parliaments are taken into account. This may have had a particular impact on forecast errors and bias, notably for year-ahead forecasts, which are mostly based on a no-policy- 17
change assumption. The government budget balance projections for the EU and euro area aggregates appear as unbiased for both the current year and the year-ahead forecasts. For the 2000-2017 period, the panel regression shows no bias for general government balance forecasts. At the Member State level, forecasts for Greece, Italy and Portugal exhibit positive, while forecast for Luxembourg, Austria, Denmark and Sweden exhibit negative biases for both horizons. Depending on the country and the year, a negative average error for the government balance can either mean that the surplus was underestimated or that the deficit was overestimated. For Greece, Italy and Portugal, the positive bias for that period appear to refer to an underestimation of the deficit. 4.2. ARE THE PROJECTION ERRORS PERSISTENT? If forecasters repeat the same mistakes (or compensate past mistakes by subsequent errors of the opposite sign), forecast errors will be positively (negatively) autocorrelated. In this study, we employ the Ljung-Box test for testing serial correlation in errors up to two lags. 10 An estimate is considered statistically significant when the corresponding p-value is below 0.05. The null hypothesis of absence of autocorrelation in forecast errors is then rejected at the 5% level. 4.2.1. Gross Domestic Product Between 2000 and 2017, there is no evidence of serial correlation in GDP growth forecast errors for the EU and the euro area either for the current year or year-ahead forecasts. At the Member State level, forecasts for twelve countries show statistically significant persistence in errors at the first lag of current-year forecasts. Positive serial correlation is a typical pattern when only three of these twelve countries have a negative autocorrelation coefficient. Forecasters typically tend to smooth forecasts and do not adequately capture upswings and downswings of the business cycles, as shown below. Adding another lag, the number of countries with significantly persistent errors drops to eight. For three of these countries (Bulgaria, Hungary and Poland), the sample is shorter as they entered the EU in and after 2004 hence the test may not be powerful enough, and the results should be interpreted with caution. In order to assess the robustness of the results, the test was run on the entire sample. The results (see Table A.5 in Appendix) reveal that serial correlation in errors is a systematic feature of the current-year forecasts for Greece, Spain, Italy, Austria and Sweden. For one-year-ahead forecasts, there is no evidence of serial correlation in errors for the euro area and the EU in the sample period starting in 2000. At the Member States level, the number of countries with statistically significant autocorrelated one-year-ahead forecast errors is lower than in the case of current year forecasts. The test identifies serial correlation at both lag lengths (one lag and two lags) only in forecasts for Spain. Furthermore, there is also some evidence of error persistence in forecasts for Greece Sweden and Estonia, and to a lesser extent for Slovakia, Bulgaria, Poland and Romania. 4.2.2. Inflation For inflation, no error persistence was identified either for the euro area or for the EU in the current year forecasts. At the Member States level, errors in the forecasts for Germany, Bulgaria and Poland are found to be serially correlated at both lag lengths (one lag and two lags). For additional nine countries, errors are found to be autocorrelated at either the first lag or first two lags. Unlike in the case of GDP growth forecasts, positive autocorrelation does not dominate in inflation forecast errors when thirteen countries have positively correlated errors (irrespective of the magnitude of autocorrelation) and fourteen countries have a negative autocorrelation coefficient at the first lag. At two lags, negative serial correlation is prevalent (22 countries). The post-2000 period was 10 See Box 4.1 in Fioramanti et al. (2016) for details. 18
characterised by increased error persistence as the strength of serial correlation falls below the level of statistical significance at the longer sample period. For one-year-ahead forecasts of inflation, there is no evidence of serial correlation in errors for the euro area and the EU in the period 2000-2017. At the Member States level, forecast errors for Malta and Slovakia are found to be serially correlated at both lag lengths (one lag and two lags). Furthermore, errors for Belgium, Ireland, Slovenia and Romania are found to be serially correlated when considering two lags. Like in the case of current-year forecast, negative serial correlation dominates in the country sample with 19 countries having negative autocorrelation coefficients. Table 4.2.: Tests for error persistence, current-year, 2000-2017 General Government GDP Inflation Balance lag 1 lag 2 lag 1 lag 2 lag 1 lag 2 Belgium -0.47*** -0.14* -0.24 -0.13 -0.01 -0.18 Germany 0.26** -0.17 -0.3** -0.33*** 0.27 0.2* Estonia 0.26 -0.37 -0.20 -0.37 0.2** -0.43* Ireland -0.17 0.27 0.34 -0.30 0.19* 0.12 Greece 0.73*** -0.29 -0.16 -0.24 0.02 -0.23 Spain 0.44** -0.39 -0.07 -0.03 0.34* -0.38*** France -0.38*** -0.03 0.01 -0.5*** 0.11 -0.24 Italy -0.44** -0.45 0.16 -0.04 -0.45*** -0.67** Cyprus 0.21 -0.09 0.42** -0.29 -0.13 -0.20 Latvia 0.29** -0.40 0.61*** 0.08 -0.3* -0.36*** Lithuania 0.08 -0.40 0.01 0.31** 0.32 0.04 Luxembourg 0.18*** -0.47** 0.08 0.01 -0.25*** -0.34 Malta 0.11 -0.19 -0.51** -0.11 0.42 -0.12 Netherlands 0.18 0.27** 0.12 -0.36* 0.15 -0.01 Austria 0.19** 0.17** -0.23* -0.17 0.05 0.26 Portugal -0.01 0.14 0.13 -0.32*** -0.32 -0.42* Slovenia 0.25 -0.07 -0.10 -0.20 -0.02 0.01 Slovakia 0.11 -0.28 -0.15 -0.23 0.41** -0.33 Finland -0.14 -0.28 -0.08 0.33 -0.02 -0.32 Euro Area -0.05 -0.14 -0.20 -0.20 0.41* -0.57** Bulgaria 0.11** 0.38** -0.33** -0.16*** 0.49 -0.36 Czech Republic 0.13 0.02 0.27 -0.15 -0.25 -0.08 Denmark 0.12 -0.4** 0.61*** -0.27* -0.03 -0.43** Hungary 0.59** -0.8*** -0.01 -0.4*** -0.03 -0.25** Poland -0.14** -0.17** 0.56*** -0.32*** -0.13* -0.01 Romania 0.28 0.15 -0.65*** -0.14 -0.32*** 0.00 Sweden -0.20 -0.28** -0.20 0.09 -0.04 -0.25 United Kingdom 0.11 -0.30 0.23 -0.06 0.31** -0.28 EU -0.17 -0.20 0.05 -0.21* 0.23 -0.54** Notes: Significance levels: (*) 0.10, (**) 0.05, (***) 0.01. The table reports the Ljung-Box test statistics. Source: EC, Eurostat, own calculations. 4.2.3. General government balance Current-year forecasts errors of the general government balance show persistence for both the euro area and the EU at two lags while serial correlation at the first lag is only mildly statistically significant (at the 10% level) for the euro area. As regards Member States, only forecasts for Italy show a serial correlation in errors at both lag lengths. Forecasts for additional nine countries show a serial correlation in errors at either one lag or two lags. The errors for Estonia, Spain and Latvia are found to be serially correlated at one lag length and mildly statistically significant at the other lag 19
You can also read